I am working on the different aspects of the query pipeling (for DNS over TCP or any stream / sequenced packet transport). I try a new tool which sends multiple queries to check which authoritative nameserver implementations support out-of-order responses with a funny result! I never got a conclusive answer: - without unexpected things like on the fly signing auth servers are fast enough to send responses back to back - TCP (slow start?) and network (I tried from my home) delays are greater than server processing delays. So is the pipelining and its head-of-line blocking avoidance idea not appicable? No, and the RFC5966bis I-D is right with its "(especially recursive)": the processing time of a recursive server depends on external servers when the response is not in the cache. With DNSSEC validating it should be easy to find cases with a big variation in delays. To finish there is a well known real world case where a client could send a flow of queries over TCP to a recursive server: the forwarder case. And more, it is not ridiculous to force the use of TCP with a forwarder... This leads to another question: how to keep the TCP connection opened with a minimal cost (i.e., not sending keep alives :-). There is an I-D about this (in another mailing list thread please).
Regards francis.dup...@fdupont.fr _______________________________________________ DNSOP mailing list DNSOP@ietf.org https://www.ietf.org/mailman/listinfo/dnsop
